fix(approvals): a readable diff when the proposed body is not JSON, and 6.4.0

.design-sync/NOTES.md

@@ -266,7 +266,7 @@ grep -o -- '--color-warning[a-z-]*:' .design-sync/.cache/compiled.css | sort -u
 - **`cfg.dtsPropsFor` is hand-maintained and nothing verifies it against the source.** It
   is the highest-value thing in the config (see the section above) and the easiest to let
   rot — a prop renamed in `src/` leaves the design agent coding against the old name.
-- **`__APP_VERSION__` in `ds-entry.tsx` is a hardcoded `"6.3.0"`.** Cosmetic (it only feeds
+- **`__APP_VERSION__` in `ds-entry.tsx` is a hardcoded `"6.4.0"`.** Cosmetic (it only feeds
   the sidebar footer), but bump it when the app's major/minor moves, or accept the drift.
 - **`src/assets/logo_eddi.png` must stay importable.** Moving it back under `public/`
   silently reintroduces a broken image in every design using the expanded `Sidebar`.

.design-sync/ds-entry.tsx

@@ -15,7 +15,7 @@
 // component untouched. The value is cosmetic chrome; it is a static string and
 // will drift from package.json (recorded in NOTES.md's re-sync risks).
 (globalThis as unknown as { __APP_VERSION__?: string }).__APP_VERSION__ ??=
-  "6.3.0";
+  "6.4.0";
 
 // ── ui/ ────────────────────────────────────────────────────────────────────
 export { AccessibleDialog } from "@/components/ui/accessible-dialog";

HANDOFF.md

@@ -1,6 +1,6 @@
 # EDDI Manager Project Handoff
 
-## Current Status (v6.3.0, plus the post-6.3.0 EDDI alignment in #197-#203)
+## Current Status (v6.4.0, plus the post-6.3.0 EDDI alignment in #197-#203)
 
 ### How a linked account is actually consumed (traced end to end, after getting it wrong once)
 
@@ -69,6 +69,11 @@ when the agent makes an outbound call; chat attachments are uploads via
 enumerate a user's remote content, so a picker would need a new backend surface.
 
 ### Completed Phases
+- **An approval diff you can read when the proposed body is not JSON** (branch: `fix/approval-diff-invalid-json`), plus the Manager version to **6.4.0**. Reported from a live pause: the operator's `updateLlm` call proposed an LLM config with one closing brace too many (`…60000}}]}`). It parsed on neither side of the approval, so the stored document was pretty-printed and key-sorted while the proposed one stayed a single raw line — every stored line red, one green line, nothing comparable. **6511 passing (411 files)**; each review fix below was mutation-checked (reverting it fails a named test).
+  - **The broken side is re-indented, never repaired** (`src/lib/reindent-json.ts`). Whitespace outside strings is the only thing it touches, so a stray bracket stays visible as its own line. Text that does not open with `{`/`[` is compared as written. Against a broken side the diff ignores indentation — its nesting is wrong from the first stray closer on, and comparing indentation reported every later line as changed — and prints the parsed side in the broken side's key order (`jsonKeyOrder`), since alphabetical sorting of one side only turned moved keys into changes.
+  - **What changed, and what did not.** Paired changed lines mark the words that differ (`<ins>`/`<del>`, with a per-pair timeout and a render budget — two long dissimilar lines cost over a second each), and the sticky legend reads "Lines: N added · N removed · N unchanged".
+  - **Two failure modes, two warnings.** A body broken *inside* the document is rejected by EDDI. A complete document followed by more text is not: Jackson's default reader stops at the end of the first value, and EDDI's `StrictConfigurationParser` only adds `FAIL_ON_UNKNOWN_PROPERTIES`, so the write can succeed with the tail silently dropped. `parseLeadingJson` tells them apart; the approval preview says which, hedges when redaction markers are present, and hides the viewer's own caveat so the warning is not repeated.
+  - **A trailing brace was a way past the capability scan.** `detectEscalationFlags` skipped any body that failed to parse, so appending `}` to a group create hid `dynamicAgents.allowCreation` while EDDI would still store it. It now scans the leading document.
 - **Connections review findings** (branch: `fix/connections-review-findings`): a second review of the connections feature against the backend's current contract and the fixes landing in parallel there. Seven items, one commit each; **6188 passing (388 files)**, +101 tests and 3 files against the 6158 recorded below. The full run on this machine shows **60 failures in 8 files that are not this branch's** — `updates.test.ts`, `update-check-card`, `update-banner`, the operator canaries and page, `use-platform-status`, `use-send-message-409-rollback` — every one a raw `fetch` (external host, `credentials: "omit"` + `signal`, or an SSE stream) that MSW did not intercept here, so the request hit a network that is not there. The same 10 `updates.test.ts` failures reproduce at the base commit `184fde17` in the same worktree; CI, where these pass, is the source of truth for them.
   - **`CALLER_SUPPLIED` is the third binding, and the Manager now knows it.** `binding` was derived from the auth type alone, so a caller-supplied document loaded into the editor was saved back as `SERVICE` on the first unrelated edit — with an empty `valueTemplate` the backend refuses on that binding. `bindingFor(authType, requested)` honours a requested binding where the type allows it (STATIC has two legal values; every other type exactly one), `legalBindings` names the pairs, and a loaded document keeps its stored binding. The STATIC form offers the choice as a two-way radio group with the doc's reasoning — authority, not convenience: an agent holding one shared key can reach everything that key can — and, for caller-supplied, replaces the header-value field with the exact `X-EDDI-Connection-Credential: <name> <value>` line the integrator must send. The wizard now builds its document through `toStoredConnection` like the editor, so a template typed before switching the binding is not sent. The badge has a third look; the MSW fixture (`conn6`, gnowbe) carries the shape.
   - **The mirror carries the backend's newer write rules**: the name grammar `^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$`, `timeoutMs` in 1..60000 (null is the default), the literal-segment rule on `valueTemplate` (a literal half over 32 characters or with a run of 12+ from the credential alphabet — `Bearer ${vault:k}` passes, `sk-live-abcdef${vault:x}` does not; checked on the literal halves only, so a long vault key *name* never trips it), and `extraAuthParams` value rules (no `${`, ≤ 512 characters, not one long token-alphabet run) plus reserved OAuth names refused stripped of case and separators. The `code_verifier` note above is corrected: the backend fixed its denylist spelling, so normalising both sides is now the same rule. The backend's own 400 still reaches the toast verbatim, and a test pins it.
@@ -455,6 +460,7 @@ enumerate a user's remote content, so a picker would need a new backend surface.
 - 112 Backend tenancy tests passing (`mvn test`)
 
 ### Last Commit Focus
+- Frontend: the approval diff against a non-JSON body, and 6.4.0 (above). Worth carrying forward: **"not JSON" is two different outcomes on the backend.** Check what EDDI's reader actually does with a malformed body before writing a warning about it — here the obvious sentence ("EDDI will reject this") was wrong for the case that matters most, the one where the write goes through.
 - Frontend: the post-6.3.0 EDDI alignment above (#197-#203). Three things worth carrying forward:
   - **A passing test hid several of these bugs.** The GDPR counter typo survived because its test asserted the label and not the value; the MSW fixture had been written to match the page rather than the server. When a fixture and a client agree, neither is evidence. Fixtures here now describe what EDDI actually sends, including an export that answers 207.
   - **Dead API surface is a defect, not spare capacity.** The first cut of #197 shipped a `headerCount` helper, a `postWithResponse` method and a header accessor with no callers anywhere — the same fault the same set of PRs was fixing in `transferOwnership`. All three were cut before merge; `ApiResponse` carries the status and nothing else. Add the accessor with its consumer, not before it.

package.json

@@ -1,7 +1,7 @@
 {
   "name": "eddi-manager",
   "private": true,
-  "version": "6.3.0",
+  "version": "6.4.0",
   "type": "module",
   "scripts": {
     "dev": "vite",